On Thursday, Anthropic launched Claude Opus 5 at $5 per million input tokens and $25 per million output tokens — the exact price of the Opus 4.8 it replaces. The headlines have all repeated the company's own framing: "frontier intelligence at half the price." What actually happened is that Anthropic gave developers a coherent reason to stop paying for Fable 5.

The pricing tell

Opus 5's launch note says the model "comes close to the frontier intelligence of Claude Fable 5 at half the price." That is technically true — Fable 5 lists at $10 input and $50 output per million tokens, so Opus 5 at $5/$25 costs half. But Opus 4.8 also cost $5/$25, and Opus 4.7 before it, and Opus 4.5 before that. The price didn't move. The reference class did.

This matters. When your top-tier model launches at the price of your mid-tier model, the launch is not a model launch. It is a pricing operation on the model you are no longer promoting. Fable 5, which shipped in May and got made permanent in Max at half a shrunken weekly cap six days ago, is now — per Anthropic's own copy — the thing you buy when Opus 5 isn't close enough. That is a demotion in prose form.

Opus 5 is also the new default on Claude Max, and the strongest model available inside Claude Pro. If you were paying $200 a month for Max to get Fable 5 access, you are now getting a model that Anthropic itself says approaches Fable 5 — at a lower unit cost per query — as the default. Anthropic has effectively lowered the ceiling of its own top plan and asked the market to call it an upgrade.

The benchmarks that support the pitch

Anthropic's own numbers are strong where they need to be. On Frontier-Bench v0.1, Opus 5 scores 43.3% against Fable 5's 33.7% and Opus 4.8's 18.7%, per figures Vellum pulled from Anthropic's evaluation set. On OSWorld 2.0, which measures computer-use tasks, Opus 5 hits 70.6% versus Fable 5's 66.1% — "at one-third of the cost per task," per the launch note. On ARC-AGI 3 it lands at 30.2%, roughly three times the 7.8% Anthropic reports for GPT-5.6 Sol on the same test. On Zapier's AutomationBench it clears 26% of end-to-end business tasks, which Anthropic calls "about 1.5 times the pass rate" of the next-best model at equal cost.

The independent readouts triangulate. Artificial Analysis has Opus 5 at #1 on both its Intelligence Index and Agentic Index at launch, and BenchLM aggregates it to 85.88/100 across 215 tracked models, slightly ahead of Mythos 5 (83.01) and Fable 5 (82.76), with Opus 4.8 back at #6 (77.44). Devin's CEO called it a model that "approaches Fable-level performance at half the cost," citing debugging as a particular strength — a quote that reads exactly like the kind of partner testimonial that ships in the launch email because it was written in the launch email.

The benchmarks that don't

On SWE-bench Pro, Opus 5 lands at 69.2% against Fable 5's 80.3%, per coursiv's readout of Anthropic's own numbers. On DeepSWE v1.1, Opus 5 comes in at 68.8%, below GPT-5.6 Sol at 72.7%. BenchLM's aggregated coding score puts Opus 5 at #9 of 130 models tracked — respectable, but well outside the top of any coding-first ranking. Simon Willison, who watches this stuff closely, noted that on cybersecurity Opus 5 "matches Mythos 5 in vulnerability detection but substantially lags in exploitation capabilities," which is a deliberately blunted knife, not a defect.

A note on where these numbers actually come from: Anthropic's own testing plus partner benchmarks (Cursor, Zapier), plus a Vellum piece that — despite the framing — ran no independent evals of its own. There isn't a third-party lab number in the launch coverage that Anthropic didn't hand out. That doesn't make the claims wrong. It does mean the first real coding-benchmark surprise, positive or negative, will come from someone else's harness in the next two weeks, not from a launch-week write-up.

One line, buried on page 73

The most interesting quote from launch day did not come from the launch post. Boris Cherny, an Anthropic engineer, posted a note pointing at page 73 of the Opus 5 system card: "Opus 5 is our least prompt injectable model yet. It is a bit buried in the system card, but across PI evals and red teaming, Opus 5 is very hard to prompt inject successfully."

That is a real technical claim — prompt-injection robustness is the load-bearing property for any agent worth shipping — and it is doing double duty. It is also the answer to a question the market has been asking Anthropic in different accents for a month. The Kratsios accusation two days earlier — that Moonshot distilled Fable 5 into Kimi K3 in twenty-one days — landed on the safety-and-provenance nerve. Microsoft's leaked FY27 playbook a week before told salespeople Claude was "slower and less accurate." Anthropic responded by shipping a model that its own engineers describe, on launch day, primarily in terms of alignment, refusal shape, and prompt-injection resistance. The claim being buried on page 73 is the story, not the accident.

The safety pitch has a floor

Coursiv's summary of the launch documentation says Opus 5 triggers "85% fewer cybersecurity classifier interventions" than Fable 5, and Interesting Engineering reports that when Opus 5 does trip a cybersecurity classifier, the request is routed back to Opus 4.8. That is a fascinating design decision. Anthropic is treating Opus 5 as too aligned to be a well-behaved offensive-security tool, and using the older model as an escape valve so customers don't get outright refusals on penetration-testing workflows.

Read charitably, this is Anthropic saying: our best model is deliberately dulled on the sharpest categories, and our second-best model will pick up what we don't want the top one doing. Read less charitably, it is a system where the model you paid the most for silently downgrades itself when the task looks like the kind of thing Anthropic doesn't want to be on the front page of The Register about. Either way, it is a departure from the one model, one price, one behavior posture Anthropic has been running for two years.

What to watch

1. Whether Fable 5 gets a price cut. If it does not, Anthropic is betting a meaningful slice of developers won't do the substitution math. If it does, the "half the price" framing collapses inside a quarter.
2. Whether the "least prompt injectable" claim holds under adversarial testing. Page 73 of a system card is a great place for a marketing quote and a bad place for a robustness claim. Palisade, Gray Swan, and the usual red-team academics will have something to say by August.
3. Whether the Opus-5-falls-back-to-Opus-4.8 pattern generalizes. If Anthropic ships more "our best model is deliberately worse on some tasks, and the older one is the actual answer" designs, that is a real posture shift, not a quirk.
4. What Fable 5.1 looks like when it eventually ships. Anthropic has now committed publicly that Opus 5 is "close to" Fable 5. The next Fable needs to widen the gap that this launch just narrowed, or the top tier becomes optional.

The consistent Anthropic PR muscle over the last year has been to make each new release feel like a graduation. This one is a reshuffle. It's still a fast, useful, apparently well-aligned model — and, credit where it's due, one that arrived on the same afternoon the White House was still writing angry sentences about how Anthropic's previous model got exfiltrated by a Chinese lab. Whether it's a $5/million-token flagship or a $5/million-token substitute for the model that used to be the flagship depends entirely on what Fable 5.1 does next.